refactor(skills): remove mandatory bootstrap forcing from using-tracedecay

[ScriptedAlchemy]

Scope

Skill-policy review of current master at b9204dcf7de8faac4ef05af367a546aa29c32e9f, prompted by the attached skill-authoring guidance and a repo-wide pass over the bundled TraceDecay skills.

This is not a request to weaken real product invariants. It targets global skill/tool forcing and ceremony that currently runs before the model has established that TraceDecay is the right mechanism for the task.

Evidence

plugin/skills/using-tracedecay/SKILL.md is a mandatory bootstrap skill rather than a narrowly routed workflow. Its description applies to essentially every code/repo-context task, and its body says:

• if there is even a 1% chance a TraceDecay tool or skill applies, it MUST be used;
• this happens before any response/action, including clarifying questions;
• the rule is "not negotiable";
• every followed skill must be announced;
• checklist items become todos;
• native Grep/Glob/Read/tests are globally subordinated to TraceDecay.

plugin/rules/tracedecay.mdc duplicates the same policy in an alwaysApply: true rule, including the 1% threshold, a mandatory-action table, and an anti-rationalization section.

docs/SKILL-ADOPTION-RESEARCH.md documents this as an intentional copy of the Superpowers adoption strategy: mandatory session-start behavior, anti-rationalization, forced announcement, and process-shaped routing. That experiment should now be treated as historical evidence rather than a permanent authoring doctrine.

The repository already has a narrower plugin/skills/discovering-tracedecay/SKILL.md for explicit capability discovery and using-the-cli for transport fallback, so the bootstrap skill is also overlapping existing responsibilities.

Problem

The bootstrap policy optimizes TraceDecay invocation rate, not task quality or routing precision.

A globally mandatory skill can make the model perform tool/skill ceremony for tasks where:

• the user already named the exact file and operation;
• a one-line local read is sufficient;
• the content is not indexed;
• a native operation is the cheaper exact primitive;
• the correct response requires clarification before touching the repository;
• no TraceDecay-specific domain knowledge is needed.

It also contaminates skill-quality evaluation: a skill that is globally forced does not have to earn selection through a discriminative description.

Required simplification

• Remove the 1%-chance / MUST / not-negotiable bootstrap forcing from distributed TraceDecay guidance.
• Remove "announce every skill" and checklist-to-todo ceremony as a global requirement.
• Delete or sharply narrow using-tracedecay to a genuine TraceDecay-specific onboarding/routing use case. Prefer merging any useful discovery material into discovering-tracedecay rather than keeping two routing skills.
• Reduce tracedecay.mdc to actual workspace invariants and concise routing hints; it must not globally require a TraceDecay call before ordinary reasoning or native tools.
• Preserve hard product constraints close to the relevant operation: do not query internal .tracedecay DBs directly, preserve exact project/store identity, use typed CLI fallback when MCP transport fails, preserve mutation safety.
• Update docs/SKILL-ADOPTION-RESEARCH.md to mark mandatory-bootstrap conclusions as an historical adoption experiment, with current policy pointing to behavioral routing/outcome evidence instead.

Do not replace this with another always-loaded meta-skill or equivalent wording under a different name.

Behavioral acceptance

Use the existing neutral evals/agent_adoption harness rather than source-string tests.

Add both positive and negative scenarios:

• Graph-native call/impact/repo-context tasks should still select the appropriate TraceDecay workflow early and reach the correct result efficiently.
• Known-file/local-edit/non-indexed-doc/simple exact-read tasks must be allowed to complete without ceremonial TraceDecay calls when they add no value.
• A clarification-required task must not make repository calls solely to satisfy bootstrap policy.
• MCP failure should still route to the CLI fallback without raw DB access.

Score task outcome, first useful action, and efficiency. Penalize unnecessary TraceDecay/skill invocations as well as missed useful invocations. Compare relevant full/no-skills/bare conditions so improved routing cannot be claimed from raw adoption rate alone.

No live agent evals were run for this audit.

[ScriptedAlchemy]

Status (2026-09-05): already implemented on Codex's unpushed branch codex/astra-takeover-sept5 (worktree /home/zack/.codex/worktrees/48a7/tracedecay), not on any pushed ref yet. Not duplicating on #707's branch. Close when that branch lands; residuals noted by the review lane: #888 removed the only 'every MCP tool taught by a skill' gate (deferred to #891), #890/#892 leave plugin/README.md 'Announce:' paragraph and shared_skill_contract_test MIN_DESCRIPTION_CHARS gate stale until #888 lands.